This week in deep learning, we bring you Introducing GPT-6 Sol and Luna and a paper on RRSI: Regularized Recursive Self-Improvement of Agent Harnesses.
You may also enjoy Introducing Claude Opus 5.5, The plunging price of thought, a paper on An Empirical Study of Harness Design for Coding Agents, and more!
As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.
Until next week!
Industry
Anthropic ships Opus 5.5 at $4/$20 per million tokens — 40% below Opus 5 — scoring 66.4% on Terminal-Bench 4.0 and generating output more than 30% faster.
Introducing GPT-6 Sol and Luna
OpenAI answers Anthropic within hours, pricing the GPT-6 series at half the 5.6 tier and claiming Sol makes about half as many factual mistakes as its predecessor.
‘Better than DeepSeek’: Xiaomi’s MiMo-V2.6-Pro debuts as the top open weights model in the world
Xiaomi’s MIT-licensed MiMo-V2.6-Pro, a 1.02-trillion-parameter MoE activating 42 billion, ties Grok 4.7 at 46 on Artificial Analysis while charging $0.435 per million input tokens.
xAI’s Grok 4.7 lifts Terminal-Bench 4.0 from 20.3% to 38.0%, but burns 125% more output tokens than 4.6, pushing cost to $3.74 per Intelligence Index task.
PrismML launches Bonsai 2 27B, a high-intelligence AI model so small it fits on consumer hardware
PrismML compresses Qwen3.8 27B into a 5.9GB ternary Apache-2.0 model retaining roughly 98.2% of capability and running at 143 tokens per second on a consumer GeForce 5090.
Vals, backed by Andreessen Horowitz, is looking to become the gold standard for AI benchmarking
a16z-backed Vals raises $40 million to sell confidential, task-based AI benchmarks, growing revenue eightfold year over year and expanding from eight staff to twenty-five.
Anthropic opens AI-powered biology research lab
Anthropic opens a Bay Area wet lab where Claude-driven robots run experiments through a Model Hardware Standard protocol developed with the Howard Hughes Medical Institute.
MLOps / LLMOps / AgentOps
AWS debuts Strands Harness, an open-source AI agent that can be deployed in any environment
An announcement about AWS’s open-source agent harness, which runs on any cloud and claims 26% better efficiency and 77% lower cost than Claude Code on identical tasks.
Agent Memory with Engram: A Practical Guide
A practical guide about placing agent memories in a prompt so caching survives, where a 3,500-token final request read all but roughly 100 tokens from cache.
Changing the game: Using agentic AI to secure infrastructure code
An engineering post about agentic pre-submit scanning across hundreds of millions of lines of infrastructure code, with triage agents hitting 92% precision and 3% false positives.
Learning
A data analysis about inference economics, finding the cost of a fixed AI performance level has fallen roughly 47% per quarter since 2023, or about 13x per year.
The future of practice: Enabling teachers to create learning interactives with generative UI
A research blog post about generating pedagogically-guarded interactive STEM simulations, where twelve US teachers rated their custom-built interactives 8/10 on average.
How UK AISI and EvalEval Are Making Benchmark Results Reproducible
An article about publishing evaluation results as standardised Evaluation Cards, covering five benchmarks across six frontier models so reported scores carry enough context to reproduce.
Pruning LLMs Like a Physicist: Block Removal as an Ising Optimization Problem
A research write-up about framing block removal as an Ising optimization, keeping Llama-3.3-70B at 76.9 MMLU against 54.0 for the baseline after halving its depth.
Jev introduces a new shape of LLM - System One, aka Decision Models
A commentary post about decision models, a class returning only category scores and confidences rather than text, priced at $0.042 per million input tokens with free output.
How Shopify built a continual learning loop with PyTorch and vLLM
An engineering case study about compressing production failures into model weights daily, beating a frontier baseline while cutting serving cost 96% and end-to-end latency about 38%.
Libraries & Code
An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.
The self-hosted developer control center for coding agents and automations.
Papers & Publications
RRSI: Regularized Recursive Self-Improvement of Agent Harnesses
An LLM agent’s capability is largely magnified by its harness, namely the prompts, control flow, tooling, memory, and context management surrounding the frozen backbone model. Recent methods increasingly automate this process by iteratively proposing and selecting component-wise edits of an agent harness, practically establishing a form of recursive self-improvement (RSI) at the agent-system level. However, such recursive evolution may overfit by memorizing the training tasks, showing large in-distribution gains that shrink or even vanish on out-of-distribution benchmarks. We introduce Regularized Recursive Self-Improvement of Agent Harnesses (RRSI), which incorporates the principles of regularizations into harness self-improvement by constraining the evolution candidate proposal and selection. The proposer operates with a temporally annealed budget, limiting how many edits a candidate can bundle, and it encourages unexplored trajectories based on evolution history. The selector is equipped with a critic and a pruner: the critic screens benchmark-specific proposals, while the pruner, removes changes that are too small, too expensive, or no longer useful. Together these constraints favor reusable agent mechanisms over benchmark-specific ones or even noises. Across eight benchmarks spanning coding, agentic workspace and engineering design tasks, RRSI gains up to 14.1 points on the split it evolves against and up to 4.7 points on the five out-of-distribution benchmarks, while producing a harness that runs on 30% fewer policy tokens than the unregularized evolution.
An Empirical Study of Harness Design for Coding Agents
Coding harnesses shape how autonomous coding agents translate model capabilities into long-horizon software-engineering performance, yet existing work typically evaluates harnesses as monolithic systems, leaving the effectiveness of individual components unclear. To enable component-level comparisons, we study this question with a lightweight coding harness whose execution loop is fixed while three components are varied: planning, action space, and context management. Across four models evaluated on SWE-Bench Verified and Terminal-Bench 2.1, we evaluate 176 matched settings spanning five context-management strategies, four context-window budgets, and targeted ablations of planning and action space. We find that: (1) Context management becomes increasingly valuable as the context-window budget tightens, with most of its benefit coming from preventing context-overflow failures. (2) Staging rule-based elision before LLM-based summarization provides the strongest overall efficiency among the context-management strategies, whereas making elided content recoverable adds machinery that models rarely use and yields no accuracy gain. (3) Planning shifts from an accuracy scaffold for weaker models to a cost saver for stronger models, with little change in accuracy. (4) Predefined tools improve performance for models with weaker bash proficiency, whereas bash-capable models can operate effectively with a bash-only interface and achieve substantially lower cost, especially on command-line-centric tasks. Trajectory-level analysis explains these effects: context management extends execution trajectories without substantially altering agent behavior, planning changes where trajectories stop, and the action space changes the granularity at which code is written. These findings inform model- and budget-aware harness design and provide a modular framework for evaluating future harness components.
Agensh: Scaling Organizational Intelligence to 1,024 Agents
A multi-agent system can reduce latency on complex tasks by executing work concurrently. Several pioneering harness frameworks support multi-agent systems. However, the scalability of current multi-agent harnesses is often constrained by a central orchestrator’s capacity to allocate tasks and coordinate workers. To address this limitation, we introduce Agensh, a scalable self-organized multi-agent harness without a central orchestrator: concurrent workers execute a multi-agent cooperation loop, continuously gathering context, claiming and self-assigning sub-tasks, taking action and sharing findings, verifying results, and merging progress in an asynchronous manner. The loop is supported by the agentic organization infrastructure comprising three components: a shared workspace holds proposed, ongoing, and completed work; a message interface lets workers communicate; and shared context retains reusable findings and work intentions. To test the scalability of Agensh, we evaluate it on the five hardest ProgramBench tasks with GPT-5.6-sol (high). Scaling from 1 to 128 agents raises the mean final test-pass rate from 19.31% to 28.78%, an approximately 49% relative improvement. Larger organizations reach comparable test-pass rates earlier. On pandoc, scaling from 1 to 1,024 agents raises the final test-pass rate from 33.89% to 55.06%. Worker trajectories further show that different forms of self-organized cooperation gradually emerges and standardizes as the organization grows. These results reveal the number of agents as a new scaling dimension for multi-agent organizations to expand the frontier of general intelligence, offering a practical solution for complex tasks under hard latency constraints or time budgets.
bonsai 2 keeping 98% of capability in 5.9gb at 143 tokens per second on a consumer gpu is the kind of efficiency gain that changes who gets to build. wonder how long until small-and-local becomes the default for most real work.
the harness matters more than the model now